AI TRANSPARENCY NOTICE

1. What you are using

Sovereign AI Cloud serves general-purpose AI models developed by third parties, without modification of their weights, on infrastructure operated by Media Holding Development OÜ in the European Union. The models generate text from the text you send. They do not browse the internet, do not remember earlier conversations unless you resend them, and do not have access to your systems. Image, audio and video generation are not offered.

2. The models and their developers

The inference runtime is llama.cpp (MIT licence). Under Regulation (EU) 2024/1689 (the EU AI Act), the obligations of providers of general-purpose AI models lie with the model developers; Media Holding does not substantially modify the models and is not their provider. Media Holding is the provider of the API service built on them, and you are the deployer of the AI system you build with it.

3. Known limitations

• Accuracy. Outputs can be wrong, invented, outdated or incomplete, and can differ between runs. They must be reviewed before being relied on.
• Numbers. Tasks involving money, prices or quantities need the model's thinking mode enabled (measured 2026-10-12); without it, arithmetic errors are more frequent.
• Languages. No served model writes publishable Estonian; Latvian is usable only from the 9B model and with editing (measured 2026-10-12). Major European languages are handled by the 9B and 12B models; quality varies by language and should be tested for your use.
• Bias. The models reflect their training data and can reproduce stereotypes or unequal treatment. Do not use outputs as the sole basis for decisions with legal or similarly significant effects on people.
• Context. Each model has a maximum context window (16,384 / 32,768 / 20,480 tokens as served); text beyond it is not seen.
• Safety. The models have built-in refusals that are imperfect in both directions: they may refuse legitimate requests and may comply with harmful ones. Your application needs its own safeguards.
• Determinism. Even at temperature 0 small differences between runs are possible because of batching and numerical effects.

4. Your obligations as deployer under Article 50 of the EU AI Act

Article 50 applies since 2 August 2026. When you build a product on the platform you must, in short:

1. Tell people they are interacting with an AI system (Article 50(1) and 50(3)) when your product converses with natural persons, unless that is obvious from the circumstances. Say it at the first interaction, in plain language.
2. Mark AI-generated text that you publish to inform the public on matters of public interest as artificially generated (Article 50(4)), unless the text underwent human review and editorial control and a person or organisation bears editorial responsibility for it.
3. Do not use the platform for emotion recognition or biometric categorisation of natural persons; these are either prohibited (Article 5) or carry notification duties you cannot meet with a text model.
4. Give your users the information required by other law — the GDPR (Articles 13 and 14), consumer law where it applies to your product, and sector rules.

5. What we provide to you

• This notice, the model table above and the measured limitations, updated when a model changes.
• Links to the developers' model cards and licences in the documentation.
• A request identifier on every response, so you can trace an output to a request.
• The usage metadata of your project (token counts, requests, models) in the cabinet.
• Machine-readable marking of text outputs (Article 50(2)) — the obligation of a provider of an AI system that generates synthetic text to mark its output in a machine-readable format. The technical approach for text (response metadata, provenance manifests) is under assessment with our lawyer and is PLANNED; until it is implemented, outputs are not marked, and you must handle the marking of published content on your side as described in Section 4.

6. Reporting problems

If you observe an output that is unlawful, harmful or systematically wrong, tell us at [CONTACT EMAIL] with the request identifier. Because request content is not stored, the identifier lets us see the metadata of the request but not its content; include the relevant text in your report if you can.

7. Changes

We update this notice when a model is added, replaced or removed, when a limitation is measured, and when the law changes. The revision date on this page records the last change.